We would now like to extend the K-theory functor from \(\Cat _{\perf }\) to \(\Pr ^{\dual }\). Characterization (1) of Theorem 1.2 suggests the following: if \(C\) is dualizable, write it as the kernel of a localization \(D \to E\) of compactly generated categories along strongly continuous functors, and define \[ K^{\cont }(C) := \fib \big (K(D^{\omega }) \to K(E^{\omega })\big ). \] If \(C\) is itself compactly generated, then \(C^{\omega } \to D^{\omega } \to E^{\omega }\) is a short exact sequence of small stable categories by the Thomason–Neeman localization theorem (Theorem 2.17), so this gives back \(K(C^{\omega })\). But it is not clear that this is well-defined in general: we would like to have a canonical way of writing \(C\) as the fiber of a map between compactly generated categories.
Such a canonical presentation is provided by characterization (3). Let \(C \in \Pr ^{\dual }\) and let \(\kappa \) be a regular cardinal which is large enough for the left adjoint \(\hat {y}\) of \(\colim \colon \Ind (C^{\kappa }) \to C\) to exist; in fact, \(\kappa = \omega _1\) always suffices (Theorem 3.20). We get a localization sequence in \(\PrLst \) with strongly continuous functors \[ C \xhookrightarrow {\hat {y}} \Ind (C^{\kappa }) \to \Ind (\Calk ^{\cont }_{\kappa }(C)), \] where the (continuous) Calkin category \(\Calk ^{\cont }_{\kappa }(C)\) is defined as the category of compact objects of the quotient \(\Ind (C^{\kappa })/\hat {y}(C)\).
Definition 1.9. Let \(F\) be a localizing invariant. Then we define \[ F^{\cont }(C) := \fib \big (F(C^{\kappa }) \to F(\Calk ^{\cont }_{\kappa }(C))\big ) \] for \(\kappa = \omega _1\). When \(F\) is K-theory, \(K^{\cont }(C)\) is called the continuous K-theory of \(C\).
For many localizing invariants, such as K-theory, the category \(C^{\omega _1}\) is invisible: it admits countable coproducts, so by an Eilenberg swindle we have \(K(C^{\omega _1}) = 0\), and hence \[ K^{\cont }(C) \simeq \Omega K(\Calk ^{\cont }_{\omega _1}(C)). \]
At first sight this looks rather crazy: in order to compute \(K^{\cont }_0(C)\), we have to compute \(K_1\) of a huge and inexplicit category. Fortunately, this is not how one should think about continuous K-theory in practice. The following theorem gives a more conceptual characterization, which also makes much more flexible computational techniques available.
Theorem 1.10 (Efimov, [Efimov (2024), Theorem 8.10]). Restriction along \(\Ind \) induces an equivalence of categories \[ \Fun _{\loc }(\Pr ^{\dual },\Sp ) \iso \Fun _{\loc }(\Cat _{\perf },\Sp ), \] with inverse given by \(F \mapsto F^{\cont }\). Moreover, \(F\) commutes with filtered colimits if and only if \(F^{\cont }\) does.
Here, on the left-hand side, a localizing invariant is a functor \(\Pr ^{\dual } \to \Sp \) that sends short exact sequences of dualizable categories, i.e. localization sequences in \(\PrLst \) whose functors are strongly continuous, to fiber sequences. The key input in the proof is that the construction \(C \mapsto \Calk ^{\cont }_{\omega _1}(C)\) sends short exact sequences of dualizable categories to short exact sequences of small categories. In particular, \(K^{\cont }\) is the unique localizing invariant of dualizable categories extending K-theory, which makes the definition independent of all choices involved. A treatment of this theorem is planned for Chapter 8; for now we use the cited external result.
Example 1.11. The continuous extension of \(\THH \) admits a very clean description: it is the categorical trace \[ \THH \colon \Pr ^{\dual } \to \Sp , \qquad C \mapsto \tr (\id _C) \in \End _{\PrLst }(\Sp ) = \Sp , \] also known as the dimension of \(C\) as a dualizable object of \(\PrLst \). That this is a localizing invariant is a formal property of traces.